Quality-related process monitoring approach based on sparse autoencoder and comprehensive KPLS

Abstract

The partial least squares (PLS) model is widely employed in quality-related process monitoring due to its ability to effectively establish a linear relationship between process and quality variables. To extend this capability to nonlinear scenarios, kernel partial least squares (KPLS) was introduced. However, the use of a single kernel function is often inadequate for fully capturing nonlinearity. In this paper, a novel method for quality-related process monitoring that integrates sparse autoencoders (SAE) with two KPLS models, termed SAE-CKPLS, is developed. The SAE is utilized to extract representative features from the process variables, after which two KPLS models are constructed to explore the relationship between these extracted features and residuals with the quality variables. Additionally, two Hotelling's $T^2$ monitoring statistics are derived from the decomposed subspaces to detect quality-related faults. The capability and effectiveness of the proposed SAE-CKPLS method are demonstrated through applications to both a hot rolling mill process and the industrial Tennessee Eastman process (TEP) benchmark, with comparative analysis against related methods.